Abnormal risk prediction method and device based on assembly knowledge graph, equipment and medium

Abstract

An abnormal risk prediction method, device and equipment based on an assembly knowledge graph, and a medium are related to the technical field of data processing. The method comprises: acquiring multi-source heterogeneous data, correlating and aligning through a unified primary key set, and splitting to generate a time-sliced data set. An assembly field ontology and an assembly knowledge graph mode layer are constructed, the assembly field ontology is instantiated using the time-sliced data set, and a snapshot sequence of the assembly knowledge graph is constructed. Based on the snapshot sequence of the assembly knowledge graph, the influence relationship between variables in the assembly process is modeled and updated to obtain an assembly process influence relationship structure. In combination with the snapshot sequence of the assembly knowledge graph and the assembly process influence relationship structure, a state representation of a next time window is predicted by encoding and embedding representation based on a historical time window, and an abnormal risk score is calculated. When the risk exceeds a threshold, an influence path is tracked, controllable decision variables are screened, candidate intervention schemes are generated and their effects are evaluated, and a decision suggestion is output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an abnormal risk prediction method and device based on an assembly knowledge graph, equipment, and medium. BACKGROUND

[0002] The assembly process of an industrial robot joint is a representative scenario of discrete manufacturing, involving multiple processes such as press-fitting and pre-tightening of various parts such as reducers and servo motors, and integrating multiple heterogeneous data such as MES and SCADA / PLC. Due to dynamic working conditions such as equipment wear and order fluctuations, the assembly process is prone to quality problems such as torque overrun and waveform abnormalities, so there is an urgent need for a technical solution that can model the assembly structure, process constraints, and abnormal evolution path, and implement pre-event risk warning and explainable decision support.

[0003] Currently, there are various technical means in the related field trying to address the above-mentioned needs. The online monitoring method based on threshold values achieves simple alarm by setting upper and lower limits of parameters, but it is difficult to capture the coupling relationship of multiple variables. The method based on machine learning or deep learning uses historical data to build a prediction model, but lacks structured expression of process mechanisms. The knowledge graph-based technology focuses on manufacturing knowledge management, but is not deeply coupled with online anomaly detection. The intelligent method for scheduling focuses on production efficiency and pays insufficient attention to micro anomalies.

[0004] These existing methods all have obvious defects. First, they fail to integrate assembly knowledge and multi-source data in a unified framework, resulting in a separation of anomaly detection and decision support. Second, the modeling of abnormal propagation path and influence relationship is insufficient, making the results difficult to explain and translate into executable interventions. In addition, the closed-loop capability from risk warning to causal decision-making is missing, relying on manual experience and being inefficient. Finally, there is a lack of fine-grained application solutions for industrial robot joint assembly lines, limiting the applicability of actual deployment. SUMMARY

[0005] The present application provides an abnormal risk prediction method, device, equipment, and medium based on an assembly knowledge graph to improve at least one of the above technical problems.

[0006] In a first aspect, the present application provides an abnormal risk prediction method based on an assembly knowledge graph, which comprises S1 to S5.

[0007] S1, acquiring multi-source heterogeneous data of an industrial site, associating and aligning the multi-source heterogeneous data through a unified primary key set, and dividing the multi-source heterogeneous data into time-slice datasets according to fixed-length time windows.

[0008] S2, construct the assembly field ontology and the assembly knowledge graph pattern layer, and instance the assembly field ontology by using the time slice data set, and construct a sequence of assembly knowledge graph snapshots rolling with time windows.

[0009] S3, model and update the influence relationship between variables in the assembly process based on the sequence of assembly knowledge graph snapshots, and obtain an assembly process influence relationship structure containing quantified influence intensity.

[0010] S4, combine the sequence of assembly knowledge graph snapshots and the assembly process influence relationship structure to perform graph structure feature coding on a monitoring node, predict a state representation of a next time window based on an embedding representation of a historical time window, calculate an abnormal risk score to identify an abnormal risk.

[0011] S5, when the abnormal risk is identified to exceed a threshold, perform influence path tracking based on the assembly process influence relationship structure, screen controllable decision variables, generate a candidate intervention scheme and evaluate the effect thereof, and output a decision suggestion.

[0012] In a second aspect, the application provides an abnormal risk prediction device based on an assembly knowledge graph, which comprises a segmentation module, an instantiation module, a modeling module, an identification module and a scheme module.

[0013] The segmentation module is used to acquire multi-source heterogeneous data of an industrial site, associate and align the multi-source heterogeneous data through a unified primary key set, and segment the multi-source heterogeneous data according to fixed-length time windows to generate a time slice data set.

[0014] The instantiation module is used to construct the assembly field ontology and the assembly knowledge graph pattern layer, and instance the assembly field ontology by using the time slice data set, and construct a sequence of assembly knowledge graph snapshots rolling with time windows.

[0015] The modeling module is used to model and update the influence relationship between variables in the assembly process based on the sequence of assembly knowledge graph snapshots, and obtain an assembly process influence relationship structure containing quantified influence intensity.

[0016] The identification module is used to combine the sequence of assembly knowledge graph snapshots and the assembly process influence relationship structure to perform graph structure feature coding on a monitoring node, predict a state representation of a next time window based on an embedding representation of a historical time window, calculate an abnormal risk score to identify an abnormal risk.

[0017] The scheme module is used to, when the abnormal risk is identified to exceed a threshold, perform influence path tracking based on the assembly process influence relationship structure, screen controllable decision variables, generate a candidate intervention scheme and evaluate the effect thereof, and output a decision suggestion.

[0018] In a third aspect, the present application provides an abnormal risk prediction device based on an assembly knowledge graph, comprising a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement an abnormal risk prediction method based on an assembly knowledge graph as described in any one of the first aspect.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium. The computer-readable storage medium comprises a stored computer program, wherein the computer-readable storage medium controls the device where the computer-readable storage medium is located to execute an abnormal risk prediction method based on an assembly knowledge graph as described in any one of the first aspect when the computer program is running.

[0020] By adopting the above technical solution, the present application can achieve the following technical effects:

[0021] The present application integrates multi-source assembly data in a unified knowledge graph framework, realizes real-time monitoring and dynamic modeling of the industrial assembly process, not only can actively predict the abnormal risk of the future time window and give early warning, but also can automatically trace the abnormal propagation path, screen controllable variables and generate an intervention scheme that takes into account quality, efficiency and cost, thereby forming an intelligent closed loop of "prediction-tracing-decision-feedback", and significantly improving the reliability, explainability and decision response efficiency of the assembly process. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 Model framework diagram of the abnormal risk prediction method based on the assembly knowledge graph.

[0023] Fig. 2 Causal decision and intervention deduction flowchart. DETAILED DESCRIPTION

[0024] Embodiment one, please refer to Figs. 1-2 The first embodiment of the present application provides an abnormal risk prediction method based on an assembly knowledge graph, which can be executed by an abnormal risk prediction device based on an assembly knowledge graph (hereinafter referred to as: prediction device). In particular, it is executed by one or more processors in the prediction device to achieve S1-S5.

[0025] S1, acquire multi-source heterogeneous data of an industrial site, associate and align the multi-source heterogeneous data through a unified primary key set, and cut into time-sliced data sets according to a fixed length of time window.

[0026] The industrial site data acquisition and integration method can be summarized as the following steps:

[0027] S11, periodically or in real time from a production execution system (MES), a supervisory control and data acquisition (SCADA) system / PLC system, a quality detection system, a device maintenance system, and a document management system, pull incremental data to form a data set, respectively.

[0028] In an industrial robot joint assembly line, the application first acquires data related to the assembly process from multiple source heterogeneous systems by an industrial field data acquisition and integration module, to provide a unified data basis for subsequent assembly knowledge graph construction, graph structure enhanced abnormal risk prediction, and influence relationship network modeling.

[0029] Specific data sources include but are not limited to:

[0030] MES system data: including order information (order number, product model, delivery date, etc.), process route and process list, process start / end time, process person in charge, workstation number, work-in-process status, etc., used to depict the execution logic of "order-joint-process-workstation".

[0031] SCADA / PLC and field data acquisition and monitoring system data: including force-displacement curve, displacement-time curve collected by servo press, torque-angle curve, speed, acceleration collected by electric screwdriver / servo drive, workstation beat count, robot current, torque, position, device state (running, standby, fault), etc., used to depict the dynamic changes of process variables and device operating state in key assembly processes.

[0032] Quality detection system data: including online function test results (such as rotation resistance, gap, noise value, etc.), offline laboratory test results (such as life test, vibration analysis), sampling inspection records and their judgment labels (OK / NG, abnormal type, etc.), used to represent the explicit performance of assembly quality results and abnormal events.

[0033] Device maintenance and alarm record data: including fault code, alarm time, treatment measures, spare parts replacement record, maintenance personnel, downtime, etc., used to describe the device health status and maintenance behavior, to provide basis for modeling the "device state-abnormal event" path in the subsequent influence relationship network.

[0034] Process and knowledge document data: including assembly job instruction, control plan, failure mode and effects analysis (FMEA), empirical adjustment rules, typical abnormal case analysis report, etc., used to supplement explicit process knowledge, empirical rules, typical failure modes and their treatment strategies, to provide a basis for the schema and rule layer of the assembly knowledge graph.

[0035] In order to realize the unified modeling of the above-mentioned multi-source data, the application performs format conversion, time alignment and primary key matching on the data at the server side. SCADA / PLC data Quality inspection data Equipment maintenance data Document and knowledge data .

[0036] Then the full set of industrial field raw data .

[0037] .

[0038] S12, format standardization and field mapping are performed on each data source, including unit unification, code mapping, abnormal value and missing value processing, to ensure that different system data can be associated based on a unified primary key set .

[0039] S13, a unified primary key set is defined , and primary key matching and association are performed on multi-source data to generate a unified intermediate table with “order-joint-assembly process-station / equipment” as the main line .

[0040] .

[0041] In the formula , the order number is represented. The joint or product identifier is represented. The process identifier is represented. The station / equipment identifier is represented. The part serial number is represented. The event timestamp is represented.

[0042] To support subsequent entity-data assembly knowledge graph instantiation, the embodiment defines a unified primary key set . Through the primary key set , primary key matching and association are performed on multi-source data to generate an intermediate unified data table.

[0043] .

[0044] In the formula , the attribute field set corresponding to the record process parameters, quality results, alarm information, etc. is the number of records; is the record index.

[0045] S14, the timestamps are uniformly aligned and resampled, and the continuous time axis is divided into fixed-length time windows, and in each window coherent corresponding records, generating time-slice datasets , for subsequent construction of graph snapshots and temporal modeling.

[0046] For subsequent abnormal risk prediction and time-windowed modeling of graph structure enhancement, the present application divides the time axis according to fixed length , constructs discrete time windows . In each time window , the corresponding intermediate data subset is extracted , and a time snapshot input for assembling the knowledge graph is constructed accordingly.

[0047] S15, through the data interface service, the unified intermediate table and the time-slice dataset of each time window are mapped into entity nodes, attribute nodes and relationship edges in the assembly knowledge graph, and written into the graph database.

[0048] Specifically, through the data interface service, the intermediate table and each time window subset are mapped into entity nodes, attribute nodes and relationship edges in the assembly knowledge graph, and written into the backend graph database, providing a unified data foundation for subsequent "assembly knowledge graph construction - abnormal risk prediction of graph structure enhancement - impact relationship network modeling - causal decision and intervention evaluation".

[0049] Through the above steps, the present application fuses the multi-source heterogeneous data originally scattered in MES, SCADA / PLC, quality detection and maintenance systems under unified key value and time axis, laying a realizable and extensible data integration foundation for future abnormal risk prediction and causal path analysis on graph structure.

[0050] S2, build the assembly field ontology and the assembly knowledge graph mode layer, instantiate the assembly field ontology using the time-slice dataset, and build a rolling assembly knowledge graph snapshot sequence with time windows. Preferably, S2 includes S21 to S26.

[0051] Based on the foregoing industrial field data acquisition and integration results, the present application constructs an assembly field ontology for "graph structure enhanced abnormal risk prediction + impact relationship network driven causal decision" around the industrial robot joint assembly business. On the basis of traditional "equipment - process - part - quality" modeling, the ontology layer reserves information such as abnormal event representation, impact path description, and decision variable constraint on concepts and relationships, so that subsequent graph structure modeling, graph neural network prediction and causal decision can work collaboratively in a unified semantic space.

[0052] S21, define the entity types of assembly domain ontology, including order, product, part, process, station, equipment, parameter, quality index, anomaly, maintenance and operator, and define the monitoring unit entity as the dynamic carrier of key process execution.

[0053] Specifically, the application defines the following core entity types in the assembly domain ontology: order, product / joint, part, process, station, equipment, parameter, quality index, anomaly, maintenance, operator, etc.

[0054] S22, define the relationship types between design entities, including composition relationship and process sequence relationship representing static structure, and event occurrence relationship and state indication relationship representing dynamic evolution, and reserve weight attribute for influence class relationship, and reserve controllability and cost attributes for process parameter and maintenance activity entities.

[0055] The application defines the following main relationship types in the ontology, which are used to explicitly express assembly structure, process constraint, abnormal evolution and intervention path.

[0056] belongsTo (belongs to): a part belongs to a joint or product.

[0057] consistsOf (consists of): a joint consists of multiple parts.

[0058] executedAt (executed at): a process is executed at a station / equipment.

[0059] hasParameter (has parameter): process parameter associated with the process, such as press-fit target force, pressure holding time, fastening torque, etc.

[0060] leadsTo (leads to): influence relationship between process or parameter setting and quality index.

[0061] indicates (indicates): indication relationship between quality index and anomaly, such as "noise value exceeds limit" indicating "revolving part matching anomaly".

[0062] occursOn (occurs on): anomaly occurs on a specific order / part / station.

[0063] handledBy (handled by): anomaly is handled by a maintenance activity.

[0064] "Follows / Precedes" refers to the sequential relationship between processes.

[0065] compatibleWith / conflictsWith (compatible with / conflict with): Process constraints and resource compatibility relationships.

[0066] To better support "anomaly risk prediction with graph structure enhancement" and "causal decision-making driven by influence relationship networks", this invention further introduces the following design points in the relationship layer.

[0067] Design Point 1 (Relationship Weights and Directional Placement): For influence-type relationships such as leadsTo and indicates, their causal direction is marked at the ontology layer, and attribute fields such as "influence strength and confidence level" are reserved. Subsequently, in the "Assembly Process Influence Relationship Modeling and Update" module, these attributes are quantified and updated in a data-driven manner, so that the directed edges in the knowledge graph have both semantic meaning and can carry numerical weights.

[0068] Design Point 2 (Decision-Related Attribute Labeling): Add attributes such as "controllable", "adjustment cost", and "safety level" to entities such as Parameter and Maintenance and their relationships. These attributes are used to filter and prioritize different intervention variables during the causal decision generation stage, achieving automatic narrowing from "all possible causes" to "executable intervention variables".

[0069] Design Point 3 (Anomaly Evolution and Path Pattern Templates): Abstract common "anomaly evolution templates" from the relationships related to Anomaly, such as "parameter drift → gradual change in quality indicators → minor anomaly → severe anomaly shutdown". Encode these templates through relationships such as indicates, leadsTo, and occurOn, so that the subsequent graph structure-enhanced anomaly risk prediction module can combine these templates to make more reasonable risk amplification and path inference when early signs are identified.

[0070] S23. Generate the schema layer of the assembly knowledge graph in the graph database, and define the node schema, edge schema, and rule set including process sequence constraints, parameter value constraints, and equipment compatibility constraints.

[0071] Generating the assembly knowledge graph schema layer: Through analysis of process documents, FMEA, expert interviews, and historical data, the entity and relationship types in the aforementioned ontology can be gradually improved, and corresponding attributes and constraint rules can be added. Based on this, this invention defines corresponding node and edge patterns in a graph database to form the schema layer of the assembly knowledge graph.

[0072] Node Schema: Define uniform identification, primary attribute set and extended attribute related to time / lot for entities such as Order, Joint, Process, Station, Equipment, Parameter, Quality Index, Anomaly, Maintenance, Operator, and Monitoring Unit, etc.

[0073] Edge Schema: Define endpoint type, directionality and constraint condition for relationships such as belongsTo, consistsOf, executedAt, hasParameter, leadsTo, indicates, occursOn, handledBy, follows / precedes, compatibleWith / conflictsWith, etc.

[0074] Rule and constraint set: Explicitly record process sequence constraint, parameter value constraint, equipment compatibility constraint and safety protection constraint in schema layer, to provide consistency check basis for subsequent candidate intervention scheme generation and data filtering in graph neural network training process.

[0075] Through the above assembly knowledge ontology modeling and knowledge graph schema layer construction, the invention pre-embeds the structure and semantic information highly related to abnormal detection and causal decision-making at the knowledge level, so that the subsequent "entity-data assembly knowledge graph instantiation and time slicing", "assembly process influence relationship modeling and updating", and "graph enhanced abnormal risk prediction and causal decision-making generation" modules can work collaboratively on the same graph structure, forming an overall technical solution that is both innovative and implementable.

[0076] S24, based on a unified primary key set Entity instantiation is performed on the time-sliced data set, specific business object nodes and monitoring unit nodes are created in the graph database, and the association relationship between nodes is established according to the schema layer.

[0077] After completing the ontology and schema layer (schema) design in the assembly field, the invention needs to map and instantiate the actual production data into the knowledge graph, forming a time-evolving "entity + data" integrated assembly knowledge graph, which provides a unified input for subsequent assembly process influence relationship modeling and graph structure enhanced abnormal risk prediction.

[0078] First, based on a unified primary key set Entity instantiation is performed on key business objects in the production process. For each order, each joint to be assembled, each part batch, each process execution, each equipment, and each quality detection, a corresponding entity node is created in the graph database, and a unique identifier (such as order number, part number, process number, equipment number, detection number, timestamp, etc.) is bound, which corresponds to the entity types defined in Section 2 、 、 、 、 , Equipment, QualityIndex, Anomaly, Maintenance, Operator, etc.

[0079] On this basis, the application can also introduce a monitoring unit (MonitoringUnit) or similar intermediate entity according to the combination of "process + station + equipment + joint / order", and regard each key process execution as an independent monitoring node for carrying process variables, quality results and anomaly risk scores within a specific time window, facilitating subsequent graph neural network based on uniform granularity for time series modeling and anomaly detection.

[0080] S25, data attribute hanging and feature extraction, hanging the state information of the production execution system to the corresponding node, extracting key inflection points, peak values, mean values and variance statistics from the continuous waveform data collected by the data acquisition and monitoring system to encode process feature vectors, encoding quality detection results and judgment labels as quality feature vectors, and hanging the above feature vectors to the corresponding monitoring unit nodes.

[0081] Secondly, the numerical and state type data in MES, SCADA / PLC, quality detection and maintenance records are hung to the above entity nodes to form an "entity-data" integrated structure. Specifically, the start / end time, process status, responsible person, etc. in MES are associated with the corresponding order, process, station, personnel nodes. The process data collected by SCADA / PLC, such as force-displacement curve, torque-angle curve, speed, acceleration, station beat, equipment state, etc. are obtained by feature extraction to get key inflection points, peak values, area, mean value, variance, etc. and are encoded as process feature vectors. The functional test results (rotary resistance, gap, noise value, etc.) output by the quality detection system, offline test results and judgment labels (OK / NG, anomaly type, etc.) are encoded as quality feature vectors. The alarm code, fault type, maintenance measure, downtime, etc. in the equipment maintenance and alarm record are encoded as maintenance and health status features.

[0082] For continuous time series data, the application uses a sliding time window or a fixed time slice method to extract process features from each window The signals within the window are statistically aggregated or embedded, for example, by extracting the mean, extrema, or frequency domain features within the window, or by obtaining low-dimensional embedding vectors through autoencoders / sequence encoders. Finally, these feature codes are uniformly represented as node-based... and time window Attribute vector for index And it is attached to the corresponding entity nodes (especially monitoring unit nodes) through relationship types such as "observedAtTime", "hasValue", and "hasStatus".

[0083] In matrix form, window The set of attribute vectors for all nodes within the node is as follows: This will subsequently be used as an input feature for graph structure encoding and time series prediction models.

[0084] S26. According to a fixed length Divide the timeline into segments, within each time window. Within, extract the set of active entity nodes. Relation sets and attribute vector set Construct an assembly knowledge graph snapshot corresponding to this time window. This creates a sequence of snapshots that scrolls over time.

[0085] To depict the dynamic evolution of the assembly process under different working conditions, this invention uses a time axis of fixed length. Divide the time into segments (e.g., 5 minutes or several assembly cycles) to construct discrete time windows. In each time window Internally, based on the intermediate data table Based on the entity instantiation results, extract the active entity nodes within the window and their relationships from the graph database to construct an assembly knowledge graph snapshot: In the formula For window The set of entities involved. This is the set of relationships between the aforementioned entities. It is a collection of attribute vectors.

[0086] It includes nodes related to the current production activity, such as Order, Joint, Part, Process, Station, Equipment, QualityIndex, Anomaly, Maintenance, and MonitoringUnit. Both static structural relations such as belongsTo, consistsOf, executedAt, hasParameter, follows / precedes, and time-varying relations related to events and states such as occursOn, handledBy.

[0087] The time snapshot On the one hand, it inherits the prior information of the assembly knowledge graph in the structural layer (assembly level, process sequence, resource constraints, abnormal evolution template, etc.), and on the other hand, it reflects the real-time / near real-time observation data in the current time window. The subsequent assembly process influence relationship modeling module estimates the influence strength of controllable variables on quality and abnormal events as input. The graph structure enhanced abnormal risk prediction module performs graph structure aware time series prediction on the sequence and calculates the future abnormal risk score of each monitoring node, realizing the unified modeling and prediction basis of the application "knowledge + data dual driving".

[0088] S3, based on the assembly knowledge graph snapshot sequence, model and update the influence relationship between variables in the assembly process, and obtain the assembly process influence relationship structure containing quantified influence strength.

[0089] After completing the instantiation and time slicing of the entity-data assembly knowledge graph, the application models and updates the influence relationship between key variables in the assembly process based on the assembly knowledge graph snapshot of each time window , and obtains the assembly process influence relationship structure that evolves with the working condition , providing a structural basis for subsequent graph structure enhanced abnormal risk prediction and causal decision-making.

[0090] First, select candidate variables and construct samples. In each time window , the application selects a set of candidate independent variable sets and a set of target variable sets based on entity types, relationship types, and process constraints in the assembly knowledge graph in the assembly knowledge graph snapshot .

[0091] The independent variable set mainly includes controllable process parameters (such as press-fit target force, pressure holding time, fastening torque, beat buffer time, etc.), device state variables (such as device running / standby / fault state, current, torque, temperature, vibration index, etc.), environmental variables, and important maintenance activity markers.

[0092] The target variable set ​Mainly includes key quality indicators (such as rotary resistance, gap, noise value, etc.), beat indicators (such as station beat, work-in-process residence time, etc.), abnormal event indicating variables (such as press-fit waveform abnormality mark, noise exceeding mark, buffer zone blockage mark, etc.).

[0093] To ensure the interpretability and physical rationality of the influence relationship modeling, the present application only considers influence modeling between variable pairs that have potential physical / technological links in the assembly knowledge graph. Specifically, for a target variable , its candidate independent variable set is derived from entities and their attributes that have leadsTo, hasParameter, indicates, occursOn, etc. relationships, and automatically excludes variable combinations that are obviously irrelevant to the process mechanism.

[0094] After the variable set is determined, the observation values of the corresponding variables in the window are extracted from the snapshot to form sample data for modeling, which is used for subsequent influence relationship estimation.

[0095] Then, sparse influence relationship (simplified linear model) is estimated. For each target variable , a simplified linear influence relationship model is used for sparse influence relationship estimation in the time window . The coefficient vector is automatically learned from the sample data by a regression algorithm with sparse constraints, and the coefficients with non-zero or absolute values exceeding a certain threshold are considered as evidence of significant influence.

[0096] .

[0097] In the formula, is the candidate independent variable set obtained by filtering the knowledge graph relationships. is the influence coefficient to be estimated, which reflects the influence direction and strength of the independent variable on the target variable in the window . is the residual term.

[0098] To eliminate the influence of the dimension and scale differences of different variables on the comparability of the coefficients, the present application normalizes / standardizes and before regression estimation (for example, zero mean, unit variance), so that can be used as a dimensionless influence strength for subsequent path contribution calculation and cross-variable comparison.

[0099] The influence coefficient to be estimated ​Instead of being manually specified, they are automatically learned from sample data by regression algorithms with sparsity constraints (e.g. regression methods with L1 regularization). Such algorithms have the property of "automatic variable selection": for important independent variables, they will get larger non-zero coefficients. For unimportant or redundant independent variables, their coefficients will be shrunk to near zero or even exactly zero.

[0100] Therefore, after the solution is completed, the present application directly takes the coefficients with non-zero values or absolute values exceeding a set threshold as "evidence of significant influence" in the current time window.

[0101] Finally, structural constraints and time evolution updates are performed.

[0102] According to the coefficient vector, a weighted directed edge is added or updated in the assembly knowledge graph to obtain the assembly process influence relationship structure in the current time window , and the influence relationship is corrected in combination with structural prior constraints and process common sense constraints.

[0103] Specifically, according to these coefficients, a weighted directed edge is added or updated in the assembly knowledge graph from to , and the edge weight can be taken as or its normalized value. Thus, a set of influence relationship edge sets are obtained in each time window , which are complementary to the structural relationship edge set to depict the quantitative influence relationship of "parameter / state→quality / abnormality" under the current working condition.

[0104] To avoid pseudo-associations that are inconsistent with the assembly mechanism, the present application introduces structural prior constraints and process common sense constraints in the above influence relationship modeling process. Structural prior constraints: only allow influence edges to be generated between variable pairs that have potential physical connections in the knowledge graph. That is, only when there is a reasonable path between and under the existing relationships in the ontology (such as leadsTo, hasParameter, indicates, occursOn, etc.), will be included ​​​And participate in coefficient estimation, other variables are directly excluded from modeling, from the source to reduce the risk of false association. Process common sense constraints: for the influence direction and strength that is obviously inconsistent with the process common sense (for example, under the premise that "excessive pressing force usually increases the risk of noise exceeding the standard", but the coefficient result appears "the greater the pressing force, the smaller the noise"), the invention can be inhibited or corrected by threshold rules, symbol constraints or manual review, so that the final influence relationship is closer to the real assembly mechanism.

[0105] Time evolution and smoothing update: as the time window rolls, the influence relationship structure will be updated continuously. The invention can smooth the coefficients estimated in the continuous window (such as moving average or decay weight), which ensures the sensitivity of the model to new conditions and avoids frequent and large changes in the influence relationship caused by short-term random fluctuations.

[0106] Through the above process, the assembly process influence relationship structure obtained by the invention in each time window not only utilizes the structure and process priori in the assembly knowledge graph, but also automatically selects key influence variables through sparse learning driven by data. Subsequently, it can be converted into weighted adjacent information in the graph neural network to enhance the active anomaly detection capability of graph structure perception. On the other hand, in the "abnormality-oriented influence path analysis and causal decision generation method", based on these weighted directed edges, controllable variables can be traced back from high-risk nodes to generate and evaluate different intervention schemes, thereby forming a causal style closed loop of "early warning-explanation-disposal suggestion".

[0107] S4, in combination with the assembly knowledge graph snapshot sequence and the assembly process influence relationship structure, graph structure feature coding is performed on the monitoring node, and the state representation of the next time window is predicted based on the embedding representation of the historical time window, the abnormal risk score is calculated to identify the abnormal risk. Preferably, S4 includes S41 to S46.

[0108] The graph-enhanced abnormal risk prediction module of the invention takes the assembly knowledge graph snapshot sequence of each time window and the assembly process influence relationship structure obtained in step S3 as the basis, maps multiple sources of assembly process data into time sequence feature representation with graph structure constraints, and realizes abnormal risk prediction and early warning of key processes / stations / orders by predicting the deviation between future state and actual state, and provides high-risk objects and their context information for the subsequent causal decision module.

[0109] S41, in combination with the assembly process influence relationship structure and the structure edge set in the assembly knowledge graph snapshot, an adjacency matrix containing process structure and influence weight is constructed and normalized.

[0110] In each time window , the assembly knowledge graph snapshot is assembled The graph structure feature encoding is performed. The window The node set in the window is The attribute vector set is is a real number. is the node feature dimension / embedding dimension. Each row in the attribute vector set corresponds to a feature vector of a node (such as a monitoring unit, a device, a process, etc.).

[0111] The assembly process influence relationship structure obtained in step S3 is combined and the relationship set The present application constructs an adjacency matrix containing the process structure and the influence weight and further normalizes to obtain . Among them, the adjacency element corresponding to the structure edge takes 1 (or takes the structure weight), and the adjacency element corresponding to the influence edge takes the weight value normalized by . is the weight of the influence relationship edge When the same pair of nodes exists simultaneously in the structure edge and the influence edge, a weighted superposition method is used for synthesis. . Among them is the structure adjacency matrix, is the influence adjacency matrix, is the fusion coefficient.

[0112] Subsequently, self-loops are added in a conventional manner and degree matrix normalization is performed to obtain which is used for graph aggregation, and the elements thereof simultaneously reflect the structural association and influence strength between nodes.

[0113] S42, using the graph feature transformation structure based on neighborhood aggregation, the node features are updated layer by layer to obtain the node representation:

[0114] .

[0115] .

[0116] In the formula, is the node representation of the i-th layer. is a learnable transformation matrix, is a nonlinear activation function (such as ReLU, etc.). is the node representation of the i-th layer. is the node representation of the first layer.

[0117] The number of graph aggregation / graph convolutional layers is: ,go through The node representation obtained after layer aggregation is denoted as .

[0118] S43. For each node corresponding to the monitored object With its recent Using the graph structure embedding sequence of each time window as input, a temporal prediction network is used to obtain the predicted embedding for the next time window. .

[0119] For each node corresponding to the monitored object (For example, "the monitoring unit for a certain order at a certain key pressing process"), can be obtained from... The corresponding row is read in the window. Encoding vector As the time window scrolls, a graph structure embedding sequence for that node can be formed. Each vector incorporates information on assembly structure relationships and influence weights.

[0120] After obtaining the graph structure embedding sequence of each monitored object, the present invention constructs a time series prediction model for each monitored object to predict the "normal state representation" of the next time window based on the recent trend of change.

[0121] Specifically, nodes Embedding in the most recent m time windows As input, using parameters Temporal prediction network (For example, sequence models based on loop structures or encoder-decoder structures) to obtain the predicted embedding for the next time window. .

[0122] .

[0123] S44. Calculate the final risk score, which includes the estimated prior risk and the residual risk. :

[0124] .

[0125] In the formula, This is a priori risk calculation based on the mean and covariance of historical normal embeddings. This refers to residual risk based on actual observations. For weight fusion.

[0126] During the model training phase, this invention selects historical normal operation data as training samples and minimizes the prediction embedding. With real embedding The mean square error between the predicted and the actual time series is minimized, so that the time series prediction network learns the typical pattern of the evolution of the node state over time under normal conditions.

[0127] In the online running phase, the application adopts a two-stage risk calculation method of "early warning + residual correction": at the end of the window The estimated risk prior is calculated based on the predicted embedding , for example, taking the mean and covariance of the historical normal embedding as the reference.

[0128] .

[0129] In the formula, T represents transposition. is the covariance matrix. is the inverse matrix of the covariance matrix. When the value of the risk prior exceeds the threshold value

[0130] , early warning is triggered and the intervention evaluation process is entered. When the real data of the window arrives, the residual risk is calculated. In the formula, L2 is the L2 norm.

[0131] .

[0132] The final risk score is defined as (or used to confirm) .

[0133] S45, when the final risk score exceeds the risk warning threshold value , it is determined that the next time window has abnormal risk.

[0134] Specifically, when the value of the risk prior exceeds the threshold value , it is determined that the corresponding process / station / order has potential abnormal risk. Through the above method, the system can give the risk prior before the next window starts, and perform online correction after the real observation arrives, avoiding relying only on post-warning. S46, the node

[0135] is identified as an abnormal position, and the attribute information of the node and its associated abnormal event entity are retrieved from the assembly knowledge graph to obtain a specific abnormal type description, and warning information containing the prediction time, abnormal position, abnormal type and risk score is generated.

[0136] ​​​​​Just knowing that "there is high risk in a certain process" is not enough to support the subsequent causal decision-making. The application further combines the semantic relationship and influence relationship structure in the assembly knowledge graph to identify the type and location of the abnormality.

[0137] On the one hand, using the semantic path of "quality indicator - abnormal event - process / station / part" in the knowledge graph, the high-risk monitoring node can be associated with the possible affected quality indicators and abnormal event types. For example, when the risk score of the "encoder press-fit monitoring unit" node continues to rise, and there is a close connection relationship with the "noise value exceeds the standard" and "rotation part cooperation abnormality" abnormal event nodes in the graph, the application can give a judgment result similar to "the noise exceeds the standard in the encoder press-fit process of a certain order, which may cause the noise to exceed the standard".

[0138] On the other hand, combining the influence relationship edge set obtained in step S3 , the application can identify the key influence path pointing to the high-risk node or its downstream quality indicator node. These information will be used as input for the next "abnormality-oriented influence path analysis and causal decision-making generation method" to trace back to controllable process parameters, equipment status or maintenance behavior along the influence path, supporting the decision-making of "which parameters should be adjusted and what intervention measures should be taken".

[0139] Through the above-mentioned "graph structure feature coding - time series prediction - risk score - abnormality semantic positioning" process, the application realizes an active abnormality detection mechanism consistent with the mechanism of the assembly process in the unified carrier of the assembly knowledge graph, and provides structured high-risk objects and influence path information for the subsequent causal decision-making module, forming an integrated technical solution of "abnormal risk prediction + causal decision-making support".

[0140] S5, when the abnormal risk is identified to exceed the threshold, performing influence path tracking based on the assembly process influence relationship structure, screening controllable decision variables, generating candidate intervention schemes and evaluating their effects, and outputting decision suggestions. Preferably, S5 includes S51 to S54.

[0141] When the graph-enhanced abnormal risk prediction module is in a time window for a certain process / station / order corresponding node gives an abnormality warning (i.e. abnormal risk score exceeds the threshold ), the causal decision-making module of the application performs influence path analysis and decision-making scheme generation in the assembly knowledge graph based on the assembly process influence relationship structure obtained in S3 , realizing a causal style closed loop from "finding abnormality" to "giving a feasible disposal suggestion scheme".

[0142] S51, taking the quality indicator node or abnormal event node with high risk as the end node, performing reverse search along the weighted directed edges on the assembly process influence relationship structure, calculating the path comprehensive contribution degree according to the influence weights of edges on the path, and screening out a candidate cause node set.

[0143] First, taking the quality indicator node or abnormal event node with high risk as the end node (such as a noise exceeding node, a press-fitting waveform abnormal node, etc.), performing reverse search along the weighted directed edges from bottom to top on the influence relationship edge set of the current or nearest time window.

[0144] In the search process, the application also considers prior information such as process order, structure level and resource constraints in the assembly knowledge graph: only under the premise of meeting the process order (follows / precedes), structure composition relationship (belongsTo / consistsOf) and resource compatibility (compatibleWith / conflictsWith) restrictions, reasonable paths from nodes such as “process parameters / equipment states / part batches / environment variables” to quality indicators / abnormal events are retained.

[0145] For a directed path from a candidate cause node to an end node , the influence weights of edges on the path are . Since has been estimated based on standardized / normalized variables, it can be regarded as a dimensionless influence strength, and the application defines the path comprehensive contribution degree as a function of the edge weights. is the weight of the influence relationship edge . In an optional embodiment, the logarithmic sum (i.e.

[0146] ) or the form of normalizing the path length is used as the path comprehensive contribution degree to reduce the sorting bias caused by the path length.

[0147] Among all the path sets that meet the constraints , the path contribution degrees are sorted, thereby obtaining a set of candidate cause nodes and the corresponding comprehensive contribution degrees , which are used for priority sorting of candidate causes. The set usually includes controllable process parameters, critical equipment states, specific part batches and sensitive environment variables, etc.

[0148] ​​S52, screening out variables that can be changed by parameter adjustment, process change or maintenance operation from the candidate cause node set to form a decision variable vector .

[0149] It should be noted that not all candidate cause nodes can be directly intervened by process or operation and maintenance measures. For this purpose, the application pre-annotates attributes such as "controllable", "adjustment cost", "safety level" in the assembly knowledge graph ontology and entity attributes. In the causal decision-making stage, the variables that can be changed by parameter adjustment, process change or maintenance operation are screened out from the candidate cause set C to form a decision variable set . According to the type of decision variable, a decision variable vector is constructed The number of decision variables.

[0150] Typical components include but are not limited to: press-fit target force, press-fit speed, pressure holding time, fastening torque, etc. Process parameters. Tact time buffer, whether to enable intermediate detection process, sampling frequency, etc. Process configuration. And whether to switch to standby equipment, whether to arrange preventive maintenance, etc. Operation and maintenance strategy.

[0151] In the subsequent decision scheme generation, the application will generate a number of candidate intervention schemes around the current setting and its adjustable range, and evaluate its impact on quality and abnormal risk under the influence relationship structure and assembly knowledge graph constraints.

[0152] S53, generate a number of candidate intervention schemes , and use the influence relationship structure and local linear approximation to estimate the change trend of the key indicators .

[0153] For the object with high abnormal risk and its neighborhood, the application constructs a number of candidate intervention schemes around the decision variable vector , which can include single parameter fine tuning, multi-parameter collaborative adjustment, local process rearrangement or temporary increase of detection process, etc. Strategy.

[0154] In the scheme evaluation stage, the application uses the influence relationship structure obtained in S3 and the local linear approximation to estimate the change trend of the key indicators under the scheme is the parameter change vector relative to the current setting. Define as the change vector of the quality index and the abnormal risk index corresponding to it, under the assumption of local small range adjustment, the following linear approximation can be used. Sensitivity matrix, whose elements are derived from influence coefficients and historical data statistics, reflects the first-order influence degree of each decision variable on each quality / risk indicator.

[0155] After obtaining the estimate of , the present application further defines a multi-objective evaluation function of the scheme .

[0156] .

[0157] In the formula, is the comprehensive abnormal risk indicator. is the production efficiency indicator. is the implementation cost. is the first weight coefficient. is the second weight coefficient. is the third weight coefficient.

[0158] The production efficiency indicator is the influence of the scheme on the beat / throughput production efficiency indicator. The implementation cost is the implementation cost or resource occupation (such as the cost brought by increasing detection procedures, switching equipment, and stop-line maintenance). The weight coefficients can be configured by the enterprise according to business preferences.

[0159] In actual application, heuristic search, local enumeration, or simple linear / integer programming methods can be used to find the minimum or near-optimal scheme in the candidate scheme set under the premise of meeting process constraints and safety constraints, as the recommended treatment suggestion strategy. The strategy takes into account the multiple objectives of “reducing abnormal risk”, “controlling the impact on the beat”, and “limiting the implementation cost”.

[0160] S54, selecting the scheme with the optimal evaluation function as the decision suggestion output.

[0161] The above suggestions are production management and operation and maintenance management decision support information, and do not constitute real-time control instructions for on-site equipment or robots. Finally, the present application maps the recommended scheme to business executable decision suggestions and task lists, including but not limited to: parameter recipe suggestions (such as recommended values and applicable ranges of target forces for pressing, pressure holding times, tightening torques, etc.) for MES / process parameter management systems. Process flow configuration suggestions (such as intermediate detection procedure activation / closure, sampling ratio adjustment, temporary addition of detection links or local process rearrangement, etc.) for production management. And equipment switching suggestions and maintenance work order suggestions for operation and maintenance management systems. The above suggestions can be pushed through the interface with upper-level business systems such as MES, process parameter management systems, maintenance management systems (such as CMMS), and the approval / adoption results or actual implementation results are recorded.

[0162] The application simultaneously and continuously collects information such as key quality indicators before and after implementation, abnormal risk scores and beat changes, compares and evaluates the predicted values of the foregoing and , and returns the evaluation results to the assembly knowledge graph and influence relationship modeling module as feedback signals, which are used to update the sensitivity matrix M, the influence relationship weight and the abnormal detection model parameters.

[0163] Through the process of "influence path tracking-decision variable screening-scheme generation and effect evaluation-suggestion pushing / approval-effect evaluation and feedback update", the application realizes a cause-effect style decision closed loop starting from a high-risk node on the unified carrier of the assembly knowledge graph, so that the "graph structure enhanced abnormal risk prediction" and the "influence relationship network driven cause-effect decision" are closely coupled, which not only guarantees the explainability of the decision suggestion, but also has the implementability and evolvability in engineering.

[0164] In addition, in order to ensure that "graph structure enhanced abnormal risk prediction + influence relationship network driven cause-effect decision" is interpretable, executable and sustainable in the field, the application designs a visualization, human-computer interaction and knowledge maintenance and feedback learning module in the software system, which displays the assembly knowledge graph, abnormal risk score, influence path and decision scheme to process engineers and scheduling engineers in a graphical manner, and uses artificial feedback to close-loop correct the knowledge and model.

[0165] Through the assembly knowledge graph browsing view, the entities such as order, joint, part, process, station / equipment, process parameter, quality indicator, abnormal event and maintenance activity and their relationships (such as belongsTo, consistsOf, executedAt, hasParameter, leadsTo, indicates, occursOn, etc.) are displayed in a graph view. Users can filter and expand the path according to the order, joint or process link, view the structural relationship of "order-joint-key process-station / equipment-quality indicator-abnormal event", and superimpose the quality results, abnormal marks and maintenance records of the current or recent time window, so as to provide a structural background for abnormal positioning and decision explanation.

[0166] Through the abnormal risk monitoring large screen, the abnormal risk scores output by the graph enhanced abnormal risk prediction module are visualized based on the time axis and the layout of the assembly line station. Thus, the risk curve or heat map of the key process / station is displayed in the time dimension, reflecting the change of risk over time. In the spatial dimension, the nodes are color-coded or size-coded according to the layout of the station , and the high-risk nodes are highlighted. It also supports filtering by order, joint or process type, which facilitates the discovery of risk concentrated areas.

[0167] When the system issues an abnormal early warning or gives a recommended intervention scheme for a certain node , the front-end interface provides a "influence path and decision explanation" view. On the assembled knowledge graph, highlight the influence path from the candidate cause node to the quality indicator / abnormal event node, label the influence strength of the key edge, and explain "where does the risk come from". For candidate schemes , show their corresponding parameter adjustment content (such as press-in force, pressure holding time, detection process activation, etc.) and the expected impact on risk, cycle, cost, etc. indicators, help engineers understand the pros and cons of different schemes. For the final recommended scheme , give a brief explanation, such as "prioritize adjusting certain parameters because of their high sensitivity to current abnormal indicators and low adjustment cost", to enhance the explainability and auditability of decision suggestions.

[0168] The abnormal risk prediction method of the embodiment supports engineers to maintain the assembly knowledge graph and model output through the interface. In terms of knowledge maintenance, engineers can add / modify entities and relationships (such as adding failure modes, updating process constraints, maintaining device compatibility relationships, etc.), which are written into the graph database after verification, used for subsequent influence relationship modeling and decision calculation. In terms of feedback labeling, engineers can mark "acceptance / denial / partial adoption" and other labels on the abnormal judgment and recommended scheme given by the system, and supplement the actual cause and final treatment measures to form structured feedback samples.

[0169] In terms of parameter updating, the server side updates the influence edge weight, abnormal risk threshold, sensitivity matrix, etc. periodically according to the cumulative feedback. For example, the weight of a certain influence edge can be smoothed by simple weighted smoothing.

[0170] .

[0171] wherein is the score calculated based on user feedback, is the learning rate. represents the old.

[0172] Through the above visualization, human-computer interaction and feedback learning mechanism, the invention can continuously deposit the on-site experience of engineers into the assembly knowledge graph and abnormal detection / causal decision model, gradually improve the accuracy of abnormal detection and the rationality of decision-making, and form a sustainable evolution of intelligent assembly process monitoring and decision support system.

[0173] The application is designed around the integrated process of "predicting the future - deducing the effect of treatment recommendations - closed-loop updating". The overall architecture includes an industrial field data acquisition and integration layer, an assembly knowledge graph unified representation layer, an abnormal risk prediction layer enhanced by graph structure, an assembly process influence relationship structure construction and updating layer, and a causal decision and intervention evaluation layer. The above layers are coupled with the assembly knowledge graph as a unified carrier, which is composed of schema, instantiated entity-data assembly graph, and graph snapshot sequence formed by rolling time window, so as to ensure that the knowledge semantics, entity alignment and time evolution are consistently represented in the same semantic space. The output of the application is a decision recommendation and task list for production management and operation management, which can be pushed through the upper business system interface and implemented after being approved by authorized personnel, without directly controlling the field control layer.

[0174] The principle of unified representation of assembly knowledge and data is as follows: first, based on the ontology modeling method, the order, product structure, parts, process, station / equipment, process parameters, quality indicators, abnormal events, maintenance activities and experience rules in the joint assembly process of industrial robots are abstracted as entities and relationships, and an assembly knowledge graph with assembly level, process sequence, resource constraint and typical abnormal mode is constructed. Then, the multi-source time series data such as MES order and job record, SCADA / PLC process parameter, online / offline quality detection result, equipment alarm and maintenance record are aligned and connected with the entity / relationship nodes in the knowledge graph through order number, part number, process number and time stamp, forming an integrated assembly graph with time evolution. The unified representation not only retains the structural prior of the assembly process, but also provides a unified input for subsequent local subgraph extraction, time slice modeling and abnormal propagation analysis.

[0175] The principle of abnormal risk prediction enhanced by graph structure is as follows: in view of the difficulty of single time series in describing the coupling relationship between processes, the application takes the snapshot sequence of assembly knowledge graph as input, encodes the local assembly network in the window by graph structure and obtains the state embedding of the monitored node . On this basis, a time series prediction network is constructed to output the predicted embedding of the next window . In online operation, the application can generate "prior risk priori" based on the distance between and the historical normal embedding distribution at the end of the window, and trigger an early warning when the prior risk exceeds the threshold. When the real data of the window arrives, the prediction residual between the real embedding and is calculated to calibrate and confirm the risk score online, so as to realize the active detection of "upcoming abnormality" instead of relying on post-alarm only.

[0176] The principle of modeling the influence relationship network and the abnormal propagation path is: in order to promote the abnormal risk from "phenomenon" to "mechanism" level of explainable analysis, the assembly knowledge graph and its time sequence attribute are assembled, and the assembly process influence relationship network is constructed. Specifically, the time-sliced assembly diagram snapshot is taken as input, under the premise of meeting the prior constraints such as assembly sequence constraint and resource compatibility constraint, a relationship modeling method with sparsity constraint and robustness is used to estimate the influence strength of controllable process parameters, equipment state and environmental variables on quality indicators and abnormal events, and a weighted directed relationship network is formed. The network describes the "parameter-process-quality / abnormality" propagation path, which can be updated through a sliding time window as the working condition changes, and provides a structural basis for subsequent path tracking and intervention deduction in the cause-effect style.

[0177] The principle of causal decision generation and intervention effect deduction is: when the abnormal risk prediction module with enhanced graph structure gives that there is a high abnormal risk in a future process / position / order, the invention takes the high-risk node as the terminal point, searches the possible abnormal propagation path from top to bottom under the constraints of the influence relationship network and the assembly knowledge graph, forms a candidate cause node set, and selects the decision variables (such as press-fit target force, press-fit speed, pressure holding time, fastening torque, beat buffer time, equipment switching strategy, etc.) that can be changed through process adjustment or operation and maintenance measures among them. Around these decision variables, a local intervention response model is constructed to estimate the change trend of abnormal risk, key quality indicators and production rhythm under different intervention schemes. On this basis, by combining linear approximation model and heuristic search / linear optimization, a set of candidate intervention schemes are generated under the premise of meeting the process and resource constraints, and the intervention response model is used to perform forward-looking effect deduction and multi-objective comprehensive evaluation on each scheme, so as to output the causal style decision suggestion and its explanation path which can reduce the future abnormal risk and also take into account the production capacity and cost. Through the front-end visual interface, engineers can view the assembly knowledge graph structure, key influence path, abnormal risk distribution and the predicted effect of each candidate decision scheme, and interactively maintain and feedback learn the knowledge and model, so as to form a "knowledge + data double-driven" abnormal risk prediction and causal decision support closed loop.

[0178] The principle of closed-loop monitoring and self-learning updating is that, in order to improve the robustness of the system in a large-scale and dynamic production environment, the application integrates multi-level monitoring and layered reasoning mechanisms at the software system level. On the one hand, the system monitors and checks the consistency of data quality, model input distribution, abnormal risk score and decision execution result, and timely discovers problems such as sensor noise, data missing and working condition drift. On the other hand, the system writes back the key processes and results of each link of “early warning-path explanation-scheme recommendation-execution feedback” to the assembly knowledge graph, and through incremental updating of the influence relationship structure, revising the rule constraints and incrementally training or parameter calibration of the prediction model, the prediction accuracy and decision quality are gradually improved, forming a closed-loop optimization mechanism of “predicting the future-deriving disposal suggestions-adopting feedback-model updating”.

[0179] The above technical principles enable the application to tightly couple abnormal risk prediction and causal decision support on a unified assembly knowledge graph, and realize integrated intelligent assembly process monitoring and decision support from future risk prediction to intervention scheme design and effect deduction.

[0180] The embodiment proposes an assembly knowledge graph modeling method for an industrial robot joint assembly line, unifies order dimension, assembly structure dimension, process dimension, resource dimension, quality dimension and abnormal dimension into the same graph structure, and explicitly encodes assembly sequence constraints, pairing constraints and typical abnormal patterns, providing a unified semantic basis for subsequent data fusion and decision analysis.

[0181] On the basis of multi-source assembly data, the embodiment constructs an influence relationship modeling method that fuses structure priori, estimates the influence strength between key variables through a time window, and obtains an influence relationship structure that can be updated with working condition changes, laying a foundation for causal style decision analysis.

[0182] The embodiment designs a graph-enhanced abnormal risk prediction method, which uses the local neighborhood structure of the assembly knowledge graph and the time series prediction network to conduct forward-looking abnormal risk assessment on key processes / stations / orders, realizing “early warning” instead of only alarming after the abnormality occurs.

[0183] The embodiment proposes a causal decision generation mechanism based on influence path analysis, which maps abnormal risk scores to reason paths in the assembly knowledge graph, automatically selects controllable variables and constructs candidate disposal suggestion schemes. Through comprehensive evaluation of the schemes on indicators such as quality, tempo and cost, parameter adjustment suggestions and work order suggestions are output for production management and operation and maintenance management to adopt, and approval and traceability are supported.

[0184] The embodiment realizes the closed-loop optimization of "risk assessment-suggestion pushing-adopting feedback-model updating" by integrating the engineer experience into the knowledge graph and model parameter updating through human-computer interaction and feedback learning mechanism, and is beneficial to forming the sustainable supervision and prediction decision support capability in the complex discrete manufacturing scene such as joint assembly of industrial robots.

[0185] In the embodiment two, the application provides an abnormal risk prediction device based on an assembly knowledge graph, which comprises a segmentation module, an instantiation module, a modeling module, an identification module and a scheme module.

[0186] The segmentation module is used for acquiring multi-source heterogeneous data of an industrial site, associating and aligning the multi-source heterogeneous data through a unified primary key set, and segmenting the multi-source heterogeneous data according to a fixed length time window to generate a time slice data set.

[0187] The instantiation module is used for constructing an assembly field ontology and an assembly knowledge graph mode layer, instantiating the assembly field ontology by using the time slice data set, and constructing a rolling assembly knowledge graph snapshot sequence with time windows.

[0188] The modeling module is used for modeling and updating the influence relationship between variables in the assembly process based on the assembly knowledge graph snapshot sequence to obtain an assembly process influence relationship structure containing a quantitative influence strength.

[0189] The identification module is used for combining the assembly knowledge graph snapshot sequence and the assembly process influence relationship structure, performing graph structure feature coding on a monitoring node, predicting a state representation of a next time window based on an embedding representation of a historical time window, calculating an abnormal risk score to identify an abnormal risk, and outputting a decision suggestion.

[0190] The scheme module is used for performing influence path tracking based on the assembly process influence relationship structure when the abnormal risk exceeds a threshold, screening controllable decision variables, generating a candidate intervention scheme and evaluating the effect of the candidate intervention scheme, and outputting a decision suggestion.

[0191] In the embodiment three, the application provides an abnormal risk prediction device based on an assembly knowledge graph, which comprises a processor, a memory and a computer program stored in the memory. The computer program can be executed by the processor to realize the abnormal risk prediction method based on the assembly knowledge graph as described in any one of the embodiments one.

[0192] It can be understood that the prediction device can be an electronic device with computing performance such as a portable notebook computer, a desktop computer, a server, a smart phone or a tablet computer.

[0193] Embodiment four, the application provides a computer readable storage medium. The computer readable storage medium comprises a stored computer program, wherein the computer program controls the device where the computer readable storage medium is located to perform the abnormal risk prediction method based on the assembly knowledge graph as any one of the preceding embodiments.

Claims

1. An anomaly risk prediction method based on assembly knowledge graph, characterized in that, Include: S1. Acquire multi-source heterogeneous data from the industrial site, associate and align the multi-source heterogeneous data through a unified primary key set, and divide it according to a fixed-length time window to generate a time slice dataset. S2. Construct an assembly domain ontology and an assembly knowledge graph pattern layer. Instantiate the assembly domain ontology using the time slice dataset and construct an assembly knowledge graph snapshot sequence that scrolls with the time window. S3. Based on the snapshot sequence of the assembly knowledge graph, model and update the influence relationship between variables in the assembly process to obtain the influence relationship structure of the assembly process that includes the quantification of influence intensity. S4. Combining the assembly knowledge graph snapshot sequence with the assembly process influence relationship structure, the monitoring node is encoded with graph structure features, and the state representation of the next time window is predicted based on the embedding representation of the historical time window. An abnormal risk score is calculated to identify abnormal risks. S5. When an abnormal risk is identified that exceeds the threshold, the influence path is traced based on the influence relationship structure of the assembly process, controllable decision variables are screened, candidate intervention schemes are generated and their effects are evaluated, and decision suggestions are output. S3 includes: In each time window Based on the entity types, relation types, and process constraints in the assembly knowledge graph, in snapshot Select a set of candidate independent variables and a set of target variables ; For each target variable During the time window A simplified linear influence relationship model is used to estimate sparse influence relationships. In the formula, This is the set of candidate independent variables obtained by filtering from knowledge graph relationships; The influence coefficient to be estimated is reflected in the window. Internal independent variable For target variable The direction and intensity of the influence; For residual terms; The coefficient vector is automatically learned from the sample data by a regression algorithm with sparse constraints, and coefficients that are non-zero or whose absolute value exceeds a set threshold are regarded as evidence of significant influence. Based on the coefficient vector, weighted directed edges are added or updated in the assembly knowledge graph to obtain the assembly process influence relationship structure under this time window. Furthermore, the influence relationships are modified by combining structural prior constraints and process common sense constraints; Step S5 includes: Using high-risk quality indicator nodes or abnormal event nodes as endpoint nodes, a reverse search is performed along the weighted directed edges in the influence relationship structure of the assembly process. The comprehensive contribution of the path is calculated based on the influence weight of each edge on the path, and a set of candidate cause nodes is selected. Variables that can be changed through parameter adjustment, process modification, or maintenance operations are selected from the candidate cause node set to form a decision variable vector. ; Several candidate intervention plans were generated. Furthermore, by utilizing the influence relationship structure and local linear approximation to estimate the changing trends of key indicators, a multi-objective evaluation function is calculated. ; ； In the formula, For comprehensive abnormal risk indicators; As a production efficiency indicator; For implementation costs; This is the first weighting coefficient; This is the second weighting coefficient; This is the third weighting coefficient; The solution with the optimal evaluation function is selected as the decision recommendation output.

2. The anomaly risk prediction method based on assembly knowledge graph according to claim 1, characterized in that, S2 include: Define the entity type of the assembly domain ontology, and define the monitoring unit entity as the dynamic carrier of the key process execution process; Define the relationship types between design entities, as well as the event occurrence relationships and status indication relationships that represent dynamic evolution, and reserve weight attributes for influence-type relationships, and reserve controllability and cost attributes for process parameters and maintenance activity entities; Generate a schema layer for the assembly knowledge graph in the graph database, defining node schemas, edge schemas, and a set of rules including process sequence constraints, parameter value constraints, and equipment compatibility constraints; Based on unified primary key set The time-slice dataset is instantiated into entities, and specific business object nodes and monitoring unit nodes are created in the graph database. The relationships between nodes are established based on the schema layer. The system performs data attribute attachment and feature extraction, attaches the status information of the production execution system to the corresponding node, extracts key inflection points, peak values, mean values ​​and variances from the continuous waveform data collected by the data acquisition and monitoring system, and then encodes them into process feature vectors. The quality inspection results and judgment labels are encoded into quality feature vectors, and the above feature vectors are attached to the corresponding monitoring unit nodes. According to a fixed length Divide the timeline into segments, within each time window. Within, extract the set of active entity nodes. Relation sets and attribute vector set Construct an assembly knowledge graph snapshot corresponding to this time window. This creates a sequence of snapshots that scrolls over time.

3. The anomaly risk prediction method based on assembly knowledge graph according to claim 1, characterized in that, S4 include: Combining the influence relationship structure of the assembly process Construct an adjacency matrix containing process structure and influence weights from the structural edge set in the assembly knowledge graph snapshot. And perform normalization processing; By utilizing a graph feature transformation structure based on neighborhood aggregation, node features are updated layer by layer to obtain node representations: In the formula For the first Layer node representation; For learnable transformation matrices, It is a non-linear activation function; For each node corresponding to the monitored object With its recent Using the graph structure embedding sequence of each time window as input, a temporal prediction network is used to obtain the predicted embedding for the next time window. ; Calculate the final risk score, which includes the estimated prior risk and the residual risk. : ； In the formula, The prior for the predicted risk is calculated based on the mean and covariance of historical normal embeddings; For residual risk based on actual observations; For weight fusion.

4. The anomaly risk prediction method based on assembly knowledge graph according to claim 3, characterized in that, S4 also includes: When the final risk score Risk warning threshold If so, determine that there is an abnormal risk in the next time window; Node Identify the location as an anomaly and retrieve the node from the assembly knowledge graph. The attribute information and its associated abnormal event entities are parsed to obtain specific abnormal type descriptions, and early warning information including prediction time, abnormal location, abnormal type and risk score is generated.

5. The anomaly risk prediction method based on assembly knowledge graph according to claim 1, characterized in that, S1 includes: Incremental data is periodically or in real-time retrieved from the production execution system, data acquisition and monitoring system, quality inspection system, equipment maintenance system, and document management system to form data sets respectively; Standardize the format and map the fields of each data source, including unit unification, encoding mapping, and handling of outliers and missing values; Define a unified set of primary keys ,use Perform primary key matching and association on multi-source data to generate a unified intermediate table with "order - joint - assembly process - workstation / equipment" as the main line; ； In the formula Indicates the order number; Indicates a joint or product identifier; Indicates process identification; Indicates workstation / equipment identification; Indicates the component serial number; Indicates the event timestamp; The timestamps are uniformly aligned and resampled, and the continuous timeline is divided into fixed lengths. Time window, in each window Internal aggregation of corresponding records generates a time-slice dataset. ; Through the data interface service, the unified intermediate table and the time slice datasets of each time window are mapped to entity nodes, attribute nodes and relation edges in the assembly knowledge graph, and written into the graph database.

6. An anomaly risk prediction device based on assembly knowledge graph, characterized in that, An abnormal risk prediction method based on assembly knowledge graph as described in any one of claims 1 to 5; The abnormal risk prediction device includes: The segmentation module is used to acquire multi-source heterogeneous data from industrial sites, associate and align the multi-source heterogeneous data through a unified primary key set, and segment the data according to a fixed-length time window to generate a time slice dataset. The instantiation module is used to construct the assembly domain ontology and the assembly knowledge graph pattern layer. It instantiates the assembly domain ontology using the time slice dataset and constructs a snapshot sequence of the assembly knowledge graph that scrolls with the time window. The modeling module is used to model and update the influence relationships between variables in the assembly process based on the snapshot sequence of the assembly knowledge graph, and obtain the influence relationship structure of the assembly process including the quantified influence intensity. The identification module is used to combine the assembly knowledge graph snapshot sequence with the assembly process influence relationship structure to encode the graph structure features of the monitoring nodes, predict the state representation of the next time window based on the embedding representation of the historical time window, and calculate the abnormal risk score to identify abnormal risks. The solution module is used to trace the impact path based on the impact relationship structure of the assembly process when an abnormal risk is identified to exceed a threshold, screen controllable decision variables, generate candidate intervention solutions and evaluate their effects, and output decision recommendations.

7. An anomaly risk prediction device based on assembly knowledge graph, characterized in that, It includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement an anomaly risk prediction method based on an assembly knowledge graph as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform an anomaly risk prediction method based on an assembly knowledge graph as described in any one of claims 1 to 5.

Images